1. The Use and Misuse of Structural Equation Modeling (SEM) in Social Science Researches
Prof. Dr Zainudin Awang
Faculty of Economics and Management Sciences
University Sultan Zainal Abidin (UniSZA)
21300 Kuala Terengganu, Terengganu /

2. The Coverage of Presentation
1.What Is Structural Equation Modeling (SEM)??
2.Types of SEM (Difference between SEM and PLS-SEM)
3.The Use of SEM and PLS-SEM
4.The Misuse of SEM:
Violating the Required Parametric Assumptions
(Independently, Identically, Normally Distributed data)
Mistakes – Using Wrong Data Scales, Wrong Sampling Method, Heterogeneous Target Population, Biased Sampling (Non Probability)
Problems – Data Not Normally Distributed, CFA Procedure Failed, Items Redundancy, Constructs Redundancy, Fitness Indexes Not Achieved
Solution to all mistakes – Use PLS-SEM (The Misuse of SEM)
5. Q & A

3. What Is Structural Equation Modeling??
SEM is among the most advanced statistical analysis techniques that emerged in recent decades (Hair et al., 2013)
SEM is a class of multivariate techniques that combine aspects of factor analysis and regression, enabling the researchers to simultaneously examine relationships among latent constructs.
Considering the importance of analyzing latent constructs such as Consumer Perceptions, Attitudes, or Intentions and their influence on Organizational Performance measures (stock market price, profit, turnover, market share), SEM is one of the most prominent statistical analysis techniques today (Hair et al., 2013)

4. The SEM is a powerful statistical method to solve the following analyses:
To analyse the model with latent constructs
To run the Confirmatory Factor Analysis (CFA)
To analyse multiple equations simultaneously
To analyse regressions with multi-collinearity problems
To analyze the path analysis with multiple dependents
To estimate the correlation and covariance in a model
To model inter-relationships simultaneously in a model
To model and test the mediating variables in a model
To analyse and test the moderating variables in a model

5. Types of Structural Equation Modeling:
There are two approaches to estimate the relationships in a structural equation models namely PLS-SEM and SEM
(Hair et al., 2010; Hair et al., 2011; Hair et al., 2012a; Hair et al., 2013)
Each is appropriate for a different research objective, and researchers need to understand in order to apply the correct method
Researchers should focus on the characteristics and objectives that distinguish the two methods
In situations where theory is less developed, researchers should consider the use of PLS-SEM as an alternative to SEM. Once PLS-SEM is employed, the study is considered exploratory
The estimation procedure in PLS-SEM is OLS (Ordinary Least Squares) regression-based method rather than the MLE (Maximum Likelihood Estimates) for SEM

6. HISTORY OF MULTIVARIATE ANALYSIS METHOD Type of Statistical Analysis
SEM or PLS-SEM??
HISTORY OF MULTIVARIATE ANALYSIS METHOD
Time Period
Type of Statistical Analysis
1st Generation
( )
Exploratory
Confirmatory
Exploratory Factor Analysis (EFA)
Cluster Analysis
Multidimensional Scaling
ANOVA
Logistic Regression
Multiple Regression 
2nd Generation
(1990-Now)
PLS-SEM (VB-SEM)
SEM (CB-SEM)
The Software
Smart PLS
Warp PLS
Visual PLS
Spad PLS
PLS Graph
PLS Gui
AMOS (SPSS)
LISREL
SAS
EQS, MPLUS
SIMPLIS, PRELIS
SePATH

7. History of Multivariate Statistical Analysis
History of
Multivariate
Statistical
Analysis
1. Ordinary Least Square (OLS) Regression
Galileo (1805)
2. Partial Least Square SEM
PLS-SEM (PLS-Regression)
Herman Wold (1963, 1975, 1982)
3. Structural Equation Modeling (SEM)
Karl Joreskog (1966, 1967, 1969, 1970, 1973,1979), Wiley (1973), Joreskog & Sorbom (1982)
Software comparison between SEM (SPSS-AMOS) and PLS- SEM (Smart-PLS)
1. AMOS – Jim Arbuckle (1995) available in IBM-SPSS Version 20.0 and above as one of its analysis option
2. Smart-PLS – Ringle et al. (2005) available in webpage - free download Version 1.0, 2.0 – BUT Version 3.0 (400 Euro for one year)

8. The Comparison between Amos and Smart PLS in Hair et al.(2013)
Estimator
Used
Research
Objective
Statistical
Approach
Model
Assessment
SEM
MLE
Maximum Likelihood
Confirmatory
Theory Testing
Theory Comparison
Theory
Confirmation
Parametric
Regression
ANOVA
t-Test
F-Test
Fitness Indexes
Reflect
the fitness
of the model
(Table 1, 2)
PLS_SEM
OLS
Ordinary
Least
Square
Exploratory
Prediction
Development
Exploring
Non Parametric
Wilcoxon
Mann-Whitney
Kruskal Wallis
No fitness Indexes

9. Fitness Indexes Name of category Name of index Level of acceptance
Comments
Absolute_Fit Index
Chisq
P > 0.05
Sensitive to sample size >200
RMSEA
RMSEA < 0.08
Range 0.05 to 0.10 is acceptable
GFI
GFI > 0.90
GFI = is a good fit
Incremental Fit Index
AGFI
AGFI > 0.90
AGFI = is a good fit
CFI
CFI > 0.90
CFI = 0.95 is a good fit
TLI
TLI > 0.90
TLI = is a good fit
NFI
NFI > 0.90
NFI = 0.95 is a good fit
Parsimonious Fit Index
Chisq/df
Chi square/ df < 3.0
The value less than 3.0 is considered good fit.

10. The literature support for the respective fitness indexes
Name of category
Name of index
Index full name
Literature
Absolute
Fit Index
Chisq
Discrepancy Chi Square
Wheaton et al. (1977)
RMSEA
Root Mean Square of Error Approximation
Browne and Cudeck (1993)
GFI
Goodness of Fit Index
Joreskog and Sorbom (1984)
Incremental Fit Index
AGFI
Adjusted Goodness of Fit
Tanaka and Huba (1985)
CFI
Comparative Fit Index
Bentler (1990)
TLI
Tucker-Lewis Index
Bentler and Bonett (1980)
NFI
Normed Fit Index
Bollen (1989b)
Parsimonious Fit Index
Chisq/df
Normed Chi-Square
Marsh and Hocevar (1985)

11. Caution: Many researchers apply the statistical procedure without a comprehensive understanding of its basic foundations and principles. Researchers often fail in understanding of:
Conceptual background of the research problem under study, which should be grounded in theory and applied in management
Indicator-Construct misspecification design (Chin 1998; Jarvis et al., 2003; Mckanzie, 2001; Mckenzie et al., 2005)
An inappropriate use of the necessary measurement steps, which is evident in the application of CB-SEM
Inaccurate use of sample size and population under study (Baumgartner and Homburg, 1996)

12. The Role of Theory in Academic Research
Academic research is grounded in theory, which should be confirmed or rejected, or need further research.
A model should not be developed without some underlying theory (Hair et al., 2010)
SEM is strictly theory driven because of the exact construct specification in measurement and structural model as well as necessary modification of the models during the estimating procedure (Hair et al., 2010)
PLS-SEM is also based on some theoretical foundations, but its goal is to predict the behaviour of the relationship among constructs and to explore the underlying theoretical concept.

13. Limitation of PLS-SEM
The use for theory testing and theory confirmation is limited since it has no global fitness indexes to confirm
PLS-SEM parameter estimates are not optimal regarding bias and consistency (PLS-SEM Bias)
PLS-SEM is not recommended as a universal alternative to SEM
Researchers need to apply the SEM technique that best suit their research objectives, data characteristics, and model set-up (Hair et al., 2013) – (repeated again and again in the textbook)

14. The Misuse of SEM Data Characteristics
Data characteristics such as minimum sample size, non-normal data, and scale of measurement are among the most often stated reasons for applying PLS-SEM (Hair et al., 2012b; Henseler et al., 2009). While some of the arguments are consistent with the method’s capabilities, others are not.
For example small sample size is the most often abused argument associated with the use of PLS-SEM (Goodhue et al., 2012; Marcoulides and Saunders, 2006).
The result of these misrepresentations has been scepticism in general about the use of PLS-SEM (Hair et al., 2013)
Sample size in PLS-SEM is essentially build on the properties of OLS regression, researchers should revert to more differentiated rule of thumb provided by Cohen (1992) in his statistical power analysis provided that the measurement models have an acceptable quality in terms outer loadings (loadings should be above 0.7)
Alternatively, researchers can use programs such as G*Power ( duesseldorf.de/aap/project/gpower/) to carry out power analyses specific to the model setups. So there is no rule of the thumb as 20, 30, etc.

15. Validating the Measurement Model

16. After Deleting Poor Loading Items

17. Modification Indices: Modify the Model

18. Standardized Regression Weight

19. Path Coefficients (Regression Weight)

20. Output SEM (Amos) D <--- A .134 .082 1.625 .104 Not Significant C
Latent Constructs
Estimate
S.E.
C.R. P
Result D
<--- A
.134
.082
1.625
.104
Not Significant C
.513
.102
5.058
***
Significant B
.155
.124
1.258
.208

21. Measurement Model using PLS-SEM

22. Reliability and Validity of PLS-SEM
AVE CR
R Square
Cronbachs Alpha
Communality
Redundancy A
0.6894
0.8986
0.8493 B
0.6039
0.8840
0.8386 C
0.6946
0.9190
0.8893 D
0.5249
0.8632
0.567
0.8092
0.0922
Discriminant Validity A B C D
0.8303
0.5227
0.7771
0.5498
0.5882
0.8334
0.5623
0.6023
0.6991
0.7245

23. Structural Model after Bootstrapping

24. Standard Error (STERR) t-Statistics (|O/STERR|)
Output PLS-SEM
Original Sample (O)
Sample Mean (M)
Standard Error (STERR)
t-Statistics (|O/STERR|)
Results
A -> D
0.1880
0.2010
0.0908
2.0708
Significant
B -> D
0.2349
0.2338
0.1284
1.8303
Not Significant
C -> D
0.4575
0.4629
0.0911
5.0239
Output SEM (Amos)
Latent Constructs
Estimate
S.E.
C.R. P
Result D
<--- A
0.134
0.082
1.625
.104
Not Significant C
0.513
0.102
5.058
***
Significant B
0.155
0.124
1.258
.208

25. The Skewed Data – Parametric Assumption Failed
Descriptive Statistics N
Std. Deviation
Skewness
Statistic
Std. Error B1
452
.826
-1.026
.115 B2
1.030
-.738 B3
.781
-1.038 B4
.774
-1.231 B5
.743
-1.121 B6
.754
-1.343 B7
.762
-1.245 B8
.973
-1.593 B9
.939
-1.067
B10
.849
-1.295 M1
.867
-1.669 M2
.782
-1.043 M3
.855
-.910 M4
.913
-.899 M5
.829
-1.098 M6
.772
-1.163 M7
.937
-.931 M8
.882
-.865 M9
1.088
-.331
M10
.821
-.970

26. Run Using PLS Algorithm

27. Reliability and Validity
AVE
Composite Reliability
R Square
Cronbachs Alpha A B C D
Discriminant Validity A B C D

28. The Output of the Model

29. Output PLS-SEM for Non-Normal Data
Path Coefficients (Mean, STDEV, T-Values)
Output PLS-SEM for Non-Normal Data
Original Sample (O)
Sample Mean (M)
Standard Error (STERR)
t-Statistics (|O/STERR|)
A -> D
B -> D
C -> D

30. THANK YOU FOR LISTENING
Q & A Session